Figure 4. Pose estimation algorithm training workflow.

Stage One: Creation of a manually labelled training dataset. Stage Two: Using the unlabeled images from stage one, the pose estimation algorithm estimates the desired keypoint locations (joint centers). Estimated keypoint locations are then compared to the manually labelled training data from stage one, to determine the distance between the estimated keypoint and the manually labelled keypoint. The optimization method then adjusts filters within the layers of the algorithm to try to reduce this distance and new estimated keypoints are calculated. This process is repeated until improvements to the pose estimation algorithm are negligible.